Deep Learning with Disc Photos or OCT Scans in Glaucoma Detection

Objective

To determine whether a deep learning (DL) model using retinal nerve fiber layer thickness (RNFLT) maps from OCT scans can detect glaucoma, defined by functional visual field (VF) impairment, more accurately than a DL model using disc photos (DPs). A secondary objective was to assess the diagnostic performance of these DL models across demographic groups (race, sex, and ethnicity).

Design

Retrospective cohort study at a tertiary glaucoma center utilizing OCT and DP datasets collected between 2011 and 2022.

Participants

Out of the 16 936 DP and OCT image sets, patients with Cirrus OCT images with a quality score ≥6 of 10 and reliable 24-2 Humphrey VF tests (fixation loss ≤33%, false-negative rate ≤20%, false-positive rate ≤20%), taken within 30 days of OCT, were included. Disc photos were obtained within 6 months of OCT. Data were randomly selected for training and testing of the DL models.

Testing

Development of DL models utilizing either OCT RNFLT maps or DPs to detect glaucoma based on VF-defined functional impairment.

Main Outcome Measures

The primary outcome was the area under the curve (AUC) for glaucoma detection, comparing the OCT-based DL model with the DP-based model. The secondary outcome was the AUC across demographic groups.

Results

The OCT-based DL model achieved an AUC of 0.90, significantly outperforming the DP-based model (AUC = 0.86, P < 0.005) with superior performance consistent across demographic groups. The OCT and DP model accuracies varied significantly by demographic groups. For the OCT model, AUCs were 0.93, 0.92, and 0.92 for Asians, Blacks, and Whites (P < 0.005); 0.89 for women versus 0.93 for men (P = 0.005); and 0.92 for Hispanics versus 0.94 for non-Hispanics (P < 0.005). For the DP model, corresponding AUCs for race were 0.87, 0.90, and 0.82 (P < 0.005); for sex, 0.856 versus 0.862 (P < 0.005); and for Hispanics, 0.85 versus 0.79 (P < 0.005).

Conclusions

When glaucoma diagnosis was based on functional deficit, the OCT-based DL model offered greater accuracy in detecting glaucoma than the DP-based model, likely due to its use of objective and quantitative RNFLT measurements. This work supports the use of OCT-based DL models for glaucoma detection, while observed demographic disparities underscore the need for equitable datasets to ensure fair DL-driven glaucoma diagnosis across populations.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.